A Simple Nomogram Developed for Predicting HCC Metastasis Based on Micrornas

Background Owing to lack of predictive models for HCC metastasis based on the expression of miRNAs, we aimed to develop a simple model for identi�cation of HCC patients at high risk of metastasis. Methods HCC datasets with metastasis information were acquired from the Gene Expression Omnibus (GEO), and samples were randomly divided into training group (n=169) and testing group (n=72). Based on expression of miRNAs in the training group, we developed a predictive nomogram for HCC metastasis and evaluated its performance using area under the receiver operating characteristic curve (AUC), calibration curve, decision curve and clinical impact curve analysis. Results We found that the expression of miR-30c, miR-185 and miR-323 in HCC correlated with metastasis by the least absolute shrinkage and selection operator regression (LASSO) method and multivariate logistic regression. Based on these three miRNAs, we generated the nomogram for predicting metastasis in the training group (AUC 0.869 [95% CI .813-0.925], sensitivity 80.5%, speci�city 78.9%); in testing group (0.821 [0.770-0.872], 48.5%, 92.3%). The calibration curve showed a good agreement between actual observation and prediction by nomogram. The nomogram represented high clinical net bene�ts using decision curve and clinical impact curve analysis. Moreover, total scores calculated by nomogram were higher in dead patients than that in alive patients. In addition, the predicted target genes of these 3 miRNAs correlated with tumor metastasis by functional enrichment analysis, such as �lopodium. Our easy-to-use nomogram could assist in identifying HCC patients at high risk of metastasis, which offer valuable information for clinical treatment.


Introduction
Hepatocellular carcinoma (HCC) with several characteristics-rapid progression, poor prognosis, high invasive and resistance to anti-cancer treatments-is a leading cause of cancer-related death worldwide [1,2].Although considerable advances in HCC treatment have been achieved, such as resection, ablation and immune therapy, the grim outcome are mainly ascribed to intrahepatic metastasis or recurrence.Although surgical resection is referred to as a potentially curative treatment, almost 70% of HCC patients eventually presents recurrence after resection, and postoperative 5-year survival is less than 30%-40% [3,4].microRNAs (miRNAs), small and noncoding RNA gene products, aberrantly expressed in HCC compared with normal tissues, and detecting their expression to classify cancer subtype and predict prognosis is an alternative strategy [5].Abnormal miRNA expression was involved in metastasis of HCC [6].For instance, miR-25 overexpression could promote invasion, epithilial-mesenchymaltransition (EMT) formation by targeting Rho GDP dissociation inhibitor alpha (RhoGDI1) in HCC; silencing miR-345 and miR-638 would enhance invasion and EMT of HCC cells [7].
So far, most research mainly focused on the prognostic role of single coding or noncoding gene, several coding genes (mRNA signature) in HCC [8][9][10][11][12]: TXNDC12 could activate β-catenin, and thus trigger EMT and metastasis of HCC cells via activation of; FOXM1 [13]; a six-gene metastasis signature composed of 5 mRNAs and 1 lncRNA, or a three-gene signature composed of 3 mRNAs, could successfully predict the probability of metastasis [14]; a 17 gene signature composed of immune cytokines was a superior predictor of venous metastases [10].
To data, miRNA signature for predicting HCC metastasis was limited.One study developed 20-miRNA signature to predicted metastasis in HCC patients [4].Predictably, it is not cost-effective to detected more than 15 miRNAs.Therefore, we sought to construct a simple miRNA signature to accurately predict venous metastasis in HCC patients, which could provide some information for cancer treatment in clinical management.In this study, we identi ed three independent predictive factors-miR-30c, miR-185 and miR-323-based on these 3 miRNAs, we developed a good nomogram with high speci city and sensitivity for evaluating risk of metastasis.In addition, we predicted targets of these 3 miRNAs, and the target genes correlated with tumor metastasis by functional enrichment analysis, such as lopodium.

Materials And Methods
Metastasis related miRNAs expression levels and clinical data from GEO Expression data and clinical information of GSE6857 were download from Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/).To assure the accuracy of results, each miRNA with more than 5 missing values was deleted for further analysis.For microarray datasets, we lter out eligible records according to the following principles: (1) HCC patients were pathologically diagnosed, and some of them with metastasis; (2) the sample size was more than 30.

KM-Plot
"Kaplan-Meier plotter" (KM-Plot, http://kmplot.com/analysis/) was used to determine the prognostic role of miRNAs [15].The related miRNAs data were based on CapitalBio miRNA Array.In our study, we analyzed the correlation between the overall survival time of HCC patients and the expression of two important miRNAs (miR-30c and miR-185).

Functional enrichment analyses and Protein-Protein Interaction network analysis
As previously described, R package "clusterPro ler" was used to perform functional enrichment analyses, and Protein-Protein Interaction (PPI) networks were visualized by Cytoscape [17].Cytohubba plug-in in Cytoscape was used to analyze hub genes by calculating degrees of connectivity.The hub genes are important nodes with many interactions.

Statistical Analysis
To evaluate the relative importance of each metastasis-related miRNA, we used the least absolute shrinkage and selection operator (LASSO) regression method and multivariate logistic regression.
As previously described, Nomogram was constructed with rms package and its performance was determined by area under the receiver operating characteristic curve (AUC) and calibration (calibration plots) in R [18].All the statistical analyses were done using R (version4.0.1) other than indicated.Of all miRNAs in the GSE6857 dataset, missing values less than 0.05% of the elds were imputed by expectation-maximization method via SPSS statistical software, version 25 (SPSS, Inc., USA).A p value less than 0.05 was considered statistically signi cant.

Identi cation of metastasis related miRNAs in hepatoma carcinoma
To construct a model to evaluate the risk of metastasis, we rst obtained microarray data retrieved from Gene Expression Omnibus (GEO) database, and 241 samples (GSE=6578) were randomly divided into training group (n=169) and testing group (n=72).The Flow chart is illustrated in Figure 1.In the training group, 41 patients were diagnosed with HCC metastasis, while in the testing group, 19 were diagnosed.There was no big difference in metastasis between these two groups (p> 0.05).Then we used LASSO regression analysis to screen miRNAs highly associated with metastasis.The results of the training group with maximal AUC showed 8 potential metastasis related miRNAs, including miR-30c, miR-124a, miR-125b, miR-185, miR-206, miR-323, miR-325 and miR-1 (Figure 2A and 1B).After multivariate logistic regression analysis, only miR-30c, miR-185 and miR-323 were independent risk factors for HCC metastasis (Table 1).

Nomogram construction of hepatoma carcinoma
We developed a nomogram for predicting HCC metastasis, based on 3 independent risk factors (Figure 2C).In the training group, the nomogram performed well to distinguish HCC patients with metastasis from non-metastasis with a high AUC of 0.869 (95% CI 0.813-0.925);the sensitivity and speci city were as high as 80.49% and 78.91%, respectively (Figure 3A).The optimal cutpoint of 131.59 was calculated by Cutpoint R package, and the patients in testing group were strati ed into the high group (total points >131.59) and the low group (total points ≤131.59).Consequently, the AUC in the testing group was 0.821 (95%CI 0.770-0.872)for predicting metastasis with a sensitivity of 48.5% and speci city of 92.3% (Figure 3B, Table 2).Moreover, the calibration plot for metastasis probability confer a good agreement between the prediction by our nomogram and actual observation in the training group and testing group, respectively (Figure 3C and 2D).
We further determined clinical applicability of nomogram using decision curve analysis (DCA) and clinical impact curve analysis (CICA).The results suggested our nomogram could have an optimal clinical net bene t (Figure 3E and 2F).

Association between nomogram and survival outcome
To investigate the correlation between nomogram and survival outcome, HCC patients without follow-up were removed from further analysis in the datasets.We calculated the total points of 211 cases (Alive group, n= 152; dead group, n= 59), and observed that the dead group had higher total points compared with alive group (Figure 4A).Also, the prognostic value of miR-30c, miR-185, miR-323 were obtained using KM plotter (Figure 4B).Kaplan Meier overall survival (OS) curves displayed these results: patients with high miR-30c level had a good prognosis compared with cases with low miR-30c level (n=76, HR= 0.55, p=0.042); no obvious difference was observed between low and high miR-185 levels; unfortunately, missing expression data for miR-323 in the database resulted in an unknown prognostic role.

Functional enrichment analysis of miRNA targets
We used the multiMiR package to predict miRNA-target interactions from 8 external databases, including DIANA-microT, ElMMo, MicroCosm, miRanda, miRDB, PicTar, PITA and TargetScan.For each miRNA, If the miRNA-target interactions were predicted in 3 different database, then these targets were screened out.We then performed functional enrichment analysis.These genes were enrich in the terms of transcription coregulator activity, phosphatidylinositol binding, transcription corepressor activity, kinase regulator activity by Molecular Function (MF) enrichment analysis ; in the terms of glutamatergic synapse, organelle subcompartment, lopodium by Cellular component (CC) enrichment analysis; in the terms of cell-cell adhesion via plasma-membrane adhesion molecules, embryonic organ development by Biological Process (BP) enrichment analysis (Figure 5A-C).These genes were enrich in the terms of Transcriptional Regulation by MECP2, trans−Golgi Network Vesicle Budding by Reactome pathway analysis (Figure 5D).
The PPI network of these miRNA targets is constructed by string (Supplementary Figure 1) , and MCODE (Molecular COmplex Detection ) in Cystoscope was used to generate 3 stable subnetworks (Figure 6A-C).

Discussion
Identi cation of HCC metastasis will help clinician to make a better decision for HCC patients.In this study, we identi ed that miR-30c, miR-185, miR-323 were highly associated with metastasis.Based on these three miRNAs, we constructed a simple, useful and practical nomogram to distinguish metastasis from non-metastasis with high sensitivity and speci city; also, our nomogram possessed greatly clinical net bene t using decision curve analysis and clinical impact curve analysis.The scores of nomogram were high in dead HCC patients than that in alive patients, and patients with high miR-30c level tended to have a better prognosis.
Budhu's colleges had identi ed some metastasis-related miRNAs, including high upregulated miR-219-1, miR-207, miR-185 and miR-338 as well as down-regulated miR-34a, miR-30c and miR-148a [4].Besides miR-30c and miR-185, we also found a new metastasis-related miRNA-miR-323 [4].It has been reported that both plasma and tissue miR-323 levels were signi cantly elevated in papillary thyroid cancer patients with metastasis compared with those without metastasis.MiR-323 attenuated lung cancer cell apoptosis by targeting transmembrane protein with egf-like and 2 follistatin domain [19,20].Interestingly, on the contrary, miRNA-323 inhibited cell invasion and metastasis in pancreatic ductal adenocarcinoma by direct suppression of SMAD2 and SMAD3 [21].The reasons for the opposite role of miR-323 in metastasis might be dependent on cell context.This phenomenon was very common for miRNAs: cell-type /tissuespeci c competitive mRNAs or untranslated region-binding cofactors may have effect on the binding of miR-323 to its targets; miRNA-regulated targets in different cell types depends on the relative concentration of target mRNAs; the location of miRNA targets in the cell may alter the repressive function of miRNA.For example, miR-122-mediated-downregulation of CAT-1 was reduced owing to the relocation of CAT-1 mRNA (from cytoplasmic processing bodies (PBs) to polysomes) triggered by the cellular stress [22].
Reportedly, 20-miRNA tumor signature to predict venous metastasis of HCC, while we used only 3 miRNAs to achieve a good discrimination in identifying HCC patients at high risk of developing metastases.Moreover, application of our nomogram reached a good differentiation with high AUC values of 0.869 and 0.821 in the training group and testing group respectively.The sensitivity and speci city were as high as 80.49% and 78.91%, respectively.Although the sensitivity (48.5%) was decreased in the training group but the speci ty (92.3%) was increased.Our model was a relatively cost-effective strategy with only very few miRNAs.So far, most research mainly focused on the prognostic role of single coding or noncoding gene, several coding genes (mRNA signature) in HCC.Compared with mRNA, miRNA may be the optimal molecular indicators as the following reasons: Each miRNA is capable to affect the expression of hundreds of coding genes, and thus can affect every facet of cell, such as apoptosis, proliferation and stress response [23]; meanwhile, mature miRNAs are relatively stable and easy to detect [24].Therefore, miRNA signature can be greatly helpful in diagnosis of metastasis.In this study, we predicted targets of these 3 miRNAs, and the target genes correlated with the term of tumor metastasis by functional enrichment analysis, such as lopodium, cell-cell adhesion via plasma-membrane adhesion molecules, and transcriptional Regulation by MECP2.Filopodium, cell-cell adhesion and MECP2 play important role in metastasis [25][26][27].
There were some inherent limitations in the study.First, our study had a small sample size.We were very eager to enroll more samples to construct our model.However, dataset with su cient clinical information was very few, and strict exclusion criteria (sample size > 30) aggravated this situation.Second, external validation was lacked owing to the different sequencing platforms.Therefore, more comprehensive studies with large sample size are required to con rm performance of our nomogram in clinical management.
In conclusion, we observed 3 metastasis-related miRNAs, including miR-30c, miR-185 and miRNA-miR-323.Our nomogram is practical, easy-to-use, highly sensitive and speci c to identify HCC patients at risk of metastasis.Our ndings could assist in making a better decision about HCC treatment to improve prognosis.

Declarations Tables
Function and Reactome pathway analyses We identi ed targets of miR-30c, miR-185 and miR-323.These genes subtypes were used to perform Gene oncology (GO) analysis and Reactome pathway analysis.

Figures
Figures

Figure 1 Flow
Figure 1

Figure 2 Construction
Figure 2

Figure 3 The
Figure 3

Table 1 .
Multivariate analysis of the training group

Table 2 .
Performance of nomogram for prediction of metastasis